Zed Ships Parallel Agents. First IDE Where Two AI Agents Work at the Same Time.

Zed launched Parallel Agents on April 22. One agent refactors your backend. Another updates the frontend. A third writes tests. All running simultaneously in the same editor window. No tab switching, no separate terminals, no worktree juggling.

Nobody else does this. Cursor runs one agent at a time. Windsurf runs one agent at a time. VS Code with Copilot runs one agent at a time. Zed's Rust architecture handles native concurrency in a way that Electron-based editors structurally can't match. The performance isn't bolted on. It's the reason the editor exists.

The feature uses the open Agent Client Protocol, so you're not locked into any specific model provider. Any agent that speaks ACP can plug in. That's a deliberate architectural choice: Zed wants to be the orchestration layer, not the model layer. Let Anthropic, OpenAI, and Google compete on intelligence. Zed competes on the ability to run them all at once.

235 HN points with active discussion. The comments split between people excited about the workflow implications and people skeptical about merge conflicts when two agents edit overlapping code. Both concerns are valid. I don't know how well this handles conflicting file edits in practice. That's the hard problem. But the base capability, genuine concurrent AI work in a single IDE, is a real differentiator.

Connect this to the rest of today's findings and the picture gets interesting. If 100% of your workforce uses AI (Shopify), models are getting cheap enough to run locally (Qwen 3.6), and you know which models over-edit least (the minimal editing research), then parallel agents are the next logical step. You don't need one amazing agent. You need three decent agents working in parallel with human review at the end. That's a different workflow than what Cursor and Copilot are optimized for.

For builders evaluating IDEs: try the parallel agent workflow on a real task. Split a feature into frontend, backend, and tests. Run three agents. See if the merge is clean. If it is, you just 3x'd your throughput. If it isn't, you learned something about your codebase's coupling. Either way, worth an afternoon.